def idxs2pred_memo_all(input_idx, atlas_idx):
# objs = img_obj, mask_obj
if input_idx in memo:
input_obj = np2rs(memo[input_idx][0])
input_mask_obj = np2rs(memo[input_idx][1])
else:
input_obj, input_mask_obj = idx2obj(input_idx)
# store (img arr, mask arr) in memo
memo[input_idx] = sitk.GetArrayFromImage(
input_obj), sitk.GetArrayFromImage(input_mask_obj)
if atlas_idx in memo:
atlas_obj = np2rs(memo[atlas_idx][0])
atlas_mask_obj = np2rs(memo[atlas_idx][1])
else:
atlas_obj, atlas_mask_obj = idx2obj(atlas_idx)
memo[atlas_idx] = sitk.GetArrayFromImage(
atlas_obj), sitk.GetArrayFromImage(atlas_mask_obj)
# ground truth = input mask obj
gt_bbox = mask2bbox(sitk2np(input_mask_obj))
# predicted = atlas2pred mask obj
pred_mask_obj = atlas2pred(input_obj, atlas_obj, atlas_mask_obj)
pred_bbox = mask2bbox(sitk2np(pred_mask_obj))
return pred_bbox, gt_bbox
def idxs2pred_bbox(input_idx, atlas_idx):
# objs = img_obj, mask_obj
input_obj, input_mask_obj = idx2obj(input_idx)
atlas_obj, atlas_mask_obj = idx2obj(atlas_idx)
# ground truth = input mask obj
gt_bbox = mask2bbox(sitk2np(input_mask_obj))
# predicted = atlas2pred mask obj
pred_mask_obj = atlas2pred(input_obj, atlas_obj, atlas_mask_obj)
pred_bbox = mask2bbox(sitk2np(pred_mask_obj))
# store bbox's in memo
for idx, bbox in ((input_idx, gt_bbox),
(atlas_idx, mask2bbox(sitk2np(atlas_mask_obj)))):
if idx not in memo: memo[idx] = bbox
memo[(input_idx, atlas_idx)] = pred_bbox
return pred_bbox, gt_bbox
def get_info_df_row(folder, obj):
img_obj, mask_obj = obj
shape0, shape1, shape2 = img_obj.GetSize()
imin, imax, jmin, jmax, kmin, kmax = mask2bbox(sitk2np(mask_obj))
roi_size0, roi_size1, roi_size2 = imax - imin, jmax - jmin, kmax - kmin
return {
"folder": folder,
"shape0": shape0,
"shape1": shape1,
"shape2": shape2,
"roi_size0": roi_size0,
"roi_size1": roi_size1,
"roi_size2": roi_size2,
"imin": imin,
"imax": imax,
"jmin": jmin,
"jmax": jmax,
"kmin": kmin,
"kmax": kmax
}
def idxs2csc_bbox(input_idx, atlas_idx, pad_amts):
# pad dict
pad_cols = "pad_imin", "pad_imax", "pad_jmin", "pad_jmax", "pad_kmin", "pad_kmax"
pad_dict = dict(zip(pad_cols, pad_amts))
print("Getting obj")
# objs = img_obj, mask_obj
input_obj, input_mask_obj = idx2obj(input_idx)
atlas_obj, atlas_mask_obj = idx2obj(atlas_idx)
print("Loaded obj")
# ground truth = input mask obj
gt_mask_arr = sitk2np(input_mask_obj).astype(bool)
gt_bbox = mask2bbox(gt_mask_arr)
print("Getting pred")
# predicted = atlas2pred mask obj
pred_mask_arr = sitk2np(atlas2pred(input_obj, atlas_obj,
atlas_mask_obj)).astype(bool)
pred_bbox = mask2bbox(pred_mask_arr)
# metric
align0_dice = compute_dice_coefficient(gt_mask_arr, pred_mask_arr)
print("Starting csc")
# cascade: expand margin around pred_bbox
imin, imax, jmin, jmax, kmin, kmax = pred_bbox
imin_pad, imax_pad, jmin_pad, jmax_pad, kmin_pad, kmax_pad = pad_amts
imin, jmin, kmin = [
max(0, x - pad)
for x, pad in zip((imin, jmin, kmin), (imin_pad, jmin_pad, kmin_pad))
]
imax, jmax, kmax = [
min(x - pad, shape)
for x, pad, shape in zip((imin, jmin,
kmin), (imin_pad, jmin_pad,
kmin_pad), input_obj.GetSize())
]
# cascade: re-align sub-brain (ROI + margin) region
csc = True
try:
csc_gt_mask_arr = sitk2np(input_mask_obj[imin:imax, jmin:jmax,
kmin:kmax]).astype(bool)
csc_pred_mask_arr = sitk2np(
atlas2pred(input_obj[imin:imax, jmin:jmax, kmin:kmax],
atlas_obj[imin:imax, jmin:jmax, kmin:kmax],
atlas_mask_obj[imin:imax, jmin:jmax,
kmin:kmax])).astype(bool)
print("End csc")
print("mask2bbox")
csc_pred_bbox = mask2bbox(pred_mask_arr)
csc_gt_bbox = mask2bbox(csc_gt_mask_arr)
# metric
align1_dice = compute_dice_coefficient(csc_gt_mask_arr,
csc_pred_mask_arr)
except:
print("Error")
csc = False
csc_pred_bbox = pred_bbox
csc_gt_bbox = gt_bbox
align1_dice = align0_dice
# store error
if (input_idx, atlas_idx) in csc_errors:
csc_errors[(input_idx, atlas_idx)].append(pad_dict)
else:
csc_errors[(input_idx, atlas_idx)] = [pad_dict]
print("Return row")
# store input idx, atlas idx, pad amts, bbox delta between gt and pred
return {
"input_idx": input_idx,
"atlas_idx": atlas_idx,
"align0": align0,
"align1": align1,
"csc": csc,
**pad_dict,
**bbox2diff(csc_gt_bbox, csc_pred_bbox)
}
def atlas2pred_bbox(input_obj, atlas_obj, atlas_mask_obj):
return mask2bbox(sitk2np(atlas2pred(input_obj, atlas_obj, atlas_mask_obj)))
def idxs2csc_bbox(input_idx, atlas_idx, pad_amts):
# pad dict
pad_cols = "pad_imin", "pad_imax", "pad_jmin", "pad_jmax", "pad_kmin", "pad_kmax"
pad_dict = dict(zip(pad_cols, pad_amts))
print("Getting obj")
# objs = img_obj, mask_obj
input_obj, input_mask_obj = idx2obj(input_idx)
atlas_obj, atlas_mask_obj = idx2obj(atlas_idx)
input_arr, gt_mask_arr = (sitk2np(o) for o in (input_obj, input_mask_obj))
atlas_arr, atlas_mask_arr = (sitk2np(o)
for o in (atlas_obj, atlas_mask_obj))
del input_obj
del input_mask_obj
del atlas_obj
del atlas_mask_obj
gc.collect()
print("Loaded obj")
# ground truth = input mask obj
#gt_mask_arr = sitk2np(input_mask_obj).astype(bool)
gt_bbox = mask2bbox(gt_mask_arr)
print("Getting pred")
# predicted = atlas2pred mask obj
pred_mask_arr = atlas2pred(input_arr, atlas_arr, atlas_mask_arr)
pred_bbox = mask2bbox(pred_mask_arr)
print("Pred bbox", pred_bbox)
# metric
align0_dice = compute_dice_coefficient(gt_mask_arr, pred_mask_arr)
print("align0 dice", align0_dice)
print("Starting csc")
# cascade: expand margin around pred_bbox
imin, imax, jmin, jmax, kmin, kmax = pred_bbox
imin_pad, imax_pad, jmin_pad, jmax_pad, kmin_pad, kmax_pad = pad_amts
mins = imin, jmin, kmin
maxs = imax, jmax, jmax
min_pads = imin_pad, jmin_pad, kmin_pad
max_pads = imax_pad, jmax_pad, kmax_pad
imin, jmin, kmin = [max(0, x - pad) for x, pad in zip(mins, min_pads)]
imax, jmax, kmax = [
min(x + pad, shape)
for x, pad, shape in zip(maxs, max_pads, input_arr.shape)
]
# slice obj
print("Pred bbox", pred_bbox)
print("ROI + mrg", imin, imax, jmin, jmax, kmin, kmax)
slc = np.s_[imin:imax, jmin:jmax, kmin:kmax]
# cascade: re-align sub-brain (ROI + margin) region
csc = True
try:
print("input shape", gt_mask_arr.shape)
print("Arr[slice] shape", input_arr[slc].shape)
print("Slice ", slc)
#return
csc_gt_mask_arr = gt_mask_arr[slc]
csc_pred_mask_arr = atlas2pred(input_arr[slc], atlas_arr[slc],
atlas_mask_arr[slc])
print("End csc")
print("mask2bbox csc pred")
csc_pred_bbox = mask2bbox(csc_pred_mask_arr)
print("mask2bbox gt")
csc_gt_bbox = mask2bbox(csc_gt_mask_arr)
# metric
align1_dice = compute_dice_coefficient(csc_gt_mask_arr,
csc_pred_mask_arr)
except Exception as e:
print("Error")
print(e)
csc = False
csc_pred_bbox = pred_bbox
csc_gt_bbox = gt_bbox
align1_dice = align0_dice
# store error
if (input_idx, atlas_idx) in csc_errors:
csc_errors[(input_idx, atlas_idx)].append(pad_dict)
else:
csc_errors[(input_idx, atlas_idx)] = [pad_dict]
print("Return row")
# store input idx, atlas idx, pad amts, bbox delta between gt and pred
return {
"input_idx": input_idx,
"atlas_idx": atlas_idx,
"align0": align0_dice,
"align1": align1_dice,
"csc": csc,
**pad_dict,
**bbox2diff(csc_gt_bbox, csc_pred_bbox)
}
